Terahertz spectroscopy feature extraction method based on convex combination kernel function principal component analysis

Abstract

The present invention discloses a terahertz spectroscopy feature extraction method based on convex combination kernel function principal component analysis, and relates to the field of spectrum analysis and substance type detection technologies. The method comprises: acquiring a time domain observed signal, filtering noise information by a frequency terahertz spectrum curve obtained from a discrete fourier transformation, performing cubic spline interpolation on the spectrum curve, performing resampling by intercepting comparable data in a same frequency range and completing data normalization; and performing a convex combination kernel function mapping on a sample after completing preprocessing, performing dimension reduction on a mapped basis function matrix to finally realize terahertz spectroscopy feature extraction, and analyzing intra-cluster, inter-cluster relationships of dimension reduced data. According to the terahertz spectroscopy feature extraction method based on convex combination kernel function principal component analysis, the number and high resolution of a spectrum sample are not required, and the effect of a interference signal on the feature extraction can be fully reduced, so that the feature extraction and a low dimension representation of the terahertz frequency domain spectroscopy are realized in the case of unknown substance types and numbers, and the method has a distinct cluster effect.

Description

technical field [0001] The invention relates to a method for extracting terahertz spectral features based on principal component analysis of convex combination kernel functions, and belongs to the technical field of spectral analysis and substance type detection. Background technique [0002] In the field of spectral substance detection, scholars have been engaged in the research of infrared and Raman spectroscopy since the end of the 19th century. After decades of development, the material identification technology based on infrared absorption and scattering spectra and their peak characteristics has been relatively complete. However, the commonly used light region of the infrared frequency band, the mid-infrared frequency band, corresponds to the vibration and rotation transitions of small groups in molecules, such as triatomic and diatomic groups, and the vibration frequency of such small groups in this frequency band is the same as that in macromolecules. Other substruct...

AI Technical Summary

Problems solved by technology

However, when traditional methods are used for feature extraction of frequency-domain spectra, artificial peak calibration and primary feature matching algorithms are usually used, and the validity of each data point on the spectral curve is not judged.

Most of the manual judgment simply uses the absorption peak to mark the spectrum, and the primary feature extraction algorithm does not consider the combination of different spectral samples and algorithms

These shortcomings will bring about the problem that the feature discrimination is not obvious and the dimension of the feature vector is too low. Therefore, the traditional method has uncertainty in the feature extraction of spectral data, which will greatly increase the error of material spectral classification.

Images